Leon Weber-Genzel

@lewege.bsky.social
PostDoc in NLP at MaiNLP, CIS, LMU Munich

[Leon Weber-Genzel @weberple]

@roman_klinger @uni_bamberg_of Congrats Roman! Have a great time in Bamberg, it’s a beautiful city.

[Leon Weber-Genzel @weberple]

of the evaluation in real-world biomedical research projects that shows that the tool is actually useful to practitioners. Check out https://t.co/EPsRcXsI8J

Thanks to Fabio Barth, Leonie Lorenz, Fabian Konrath, Kirsten Huska, Jana Wolf and Ulf Leser for the great collab! (2/2)

[Leon Weber-Genzel @weberple]

Happy that a key paper from my PhD is finally out: PEDL+ (https://t.co/1RiepQ3Wcf) is a tool for extracting protein-protein and protein-chemical associations from a large portion of the biomedical literature and processing the results (who doesn't love Excel tables?). Proud (1/2)

[Barbara Plank @barbara_plank]

Trustworthiness in the Era of LLMs?

With LLMs, the traditional compartmentalized notion of language "tasks" is breaking down (dashed lines). What does it mean for trust or trustworthiness? #EMNLP2023

w/ Robert Litschko, Max (@mxmeij), Rob (@robvanderg), Leon (@weberple)

[Leon Weber-Genzel @weberple]

@n_g_laskowski I would be very interested : )

[Leon Weber-Genzel @weberple]

@tomhosking @cohere @max_nlp Super interesting work! Do you know whether factual errors rated highly during RLHF decrease the factuality of the final model? And do you think that better data tooling could help to boost the reliability of human annotators, e.g. by flagging likely factual errors?

[Barbara Plank @barbara_plank]

The ACL elections are open now - below you find my statement

[Leon Weber-Genzel @weberple]

Come and check out our poster on joint 🤖&🧐pipelines for finding annotation errors. Today from 11:40 to 12:30 at TrustNLP @ #ACL2023NLP

[MaiNLPlab @MaiNLPlab]

📝 ActiveAED: A Human in the Loop Improves Annotation Error Detection
🔍 This work shows that you can improve the performance of annotation error detection methods by using them in an active learning setup.
👥@weberple, @barbara_plank
🔗https://t.co/FIHWYdAuIV
📁Findings-Short

[Leon Weber-Genzel @weberple]

@nsaphra @eripsa Aren‘t you afraid that if future LMs/agents would have some capacity to suffer and we categorically rule it out that it would take us much too long to notice, as it did for animals? Honestly, I’m a bit.

[Leon Weber-Genzel @weberple]

I will be at ACL in Toronto. If you want to (re-)connect, reach out!

[Leon Weber-Genzel @weberple]

@_rockt @ucl @uclcs @ai_ucl @UCL_DARK Congratulations 🥳!

[Leon Weber-Genzel @weberple]

@srush_nlp Trying to do the same thing for the same reason but from PyCharm. As soon as debugging became more complex, I always went back. So, I'd also appreciate pointers to blogs/YT videos :).

[Leon Weber-Genzel @weberple]

If you want to detect labelling errors in your dataset, it might be a good idea to do it in an active learning setting for maximum performance. See our ACL 2023 Findings paper ⬇️ for how you could do that. Joint work with @barbara_plank

[MaiNLPlab @MaiNLPlab]

📝 ActiveAED: A Human in the Loop Improves Annotation Error Detection
🔍 This work shows that you can improve the performance of annotation error detection methods by using them in an active learning setup.
👥@weberple, @barbara_plank
🔗https://t.co/FIHWYdAuIV
📁Findings-Short

[Alan Akbik @alan_akbik]

Announcing 🧐 OpinionGPT 🧐, a 7B-LLaMA model fine-tuned on German-language Reddit. Our model gives opinionated answers to all your questions - in German!
Try it out here: https://t.co/5YJjHlcV4n and let us know your favorite responses :D

[Leon Weber-Genzel @weberple]

@KiddoThe2B @yoavgo Do you use Notability to provide feedback on manuscripts? If so, what's your workflow?
(Annotations with OSX preview have also disappeared for me more than once 😬)

[Leon Weber-Genzel @weberple]

Does anyone have an intuition about good hyperparameter ranges for finetuning Pythia models? @AiEleuther @BlancheMinerva

[Leon Weber-Genzel @weberple]

This is a super cool resource for studying training dynamics + scaling laws. Thanks for making it public!

[Stella Biderman @BlancheMinerva]

Have you ever wanted to do an experiment on LLMs and found that none of the existing model suites met your needs? At @AiEleuther we got tired of this happening and so designed a model suite that centers enabling scientific research as its primary goal

https://t.co/YhrwFjAGbr

[Leon Weber-Genzel @weberple]

A Christian magazine interviewed @SchmidhuberAI and @ThomasMetzinger on the role of suffering in AI research. Mostly, on the possibility of the model suffering during or after training. Very interesting read, sadly in German only. https://t.co/RpbJEtqj0Z

[Barbara Plank @barbara_plank]

📢🌄🏄 LMU München: Verwaltungsangestellte:r (Teilzeit 50%) am Lehrstuhl für KI und Computerlinguistik gesucht

Bewerbe dich bis 28.03.2023 und werde Teil vom Team @MaiNLPlab @CisLmu

Psst danke fürs sharen

[Stella Biderman @BlancheMinerva]

#ML #AI #NLProc Researchers, especially academics: I’m writing a paper about accessible finetuning of transformer models. If you’re looking to finetune a LLM, where on this list of compute access do you fall? If you don’t fall on this list, what resources are accessible to you?